Discrete Cosine Transformation and Height Functions ... - ScienceDirect

3 downloads 69 Views 853KB Size Report
1877-0509 © 2015 The Authors. Published by Elsevier B.V. ... b Department of Computer Science, University College Mangalore, Karnataka, India. ... B.H. Shekar and Bharathi Pilar / Procedia Computer Science 58 ( 2015 ) 714 – 722 might be ...
Available online at www.sciencedirect.com

ScienceDirect Procedia Computer Science 58 (2015) 714 – 722

Second International Symposium on Computer Vision and the Internet (VisionNet’15)

Discrete Cosine Transformation and Height Functions based Shape Representation and Classification B.H.Shekara , Bharathi Pilarb,∗ a Department b

of Computer Science, Mangalore University, Mangalore, Karnataka, India. Email: [email protected] Department of Computer Science, University College Mangalore, Karnataka, India. Email: [email protected]

Abstract In this paper, we propose a combined classifier model based on two dimensional discrete cosine transform (2D-DCT) and Height Functions (HF) for accurate shape representation and classification. The DCT is capable of capturing the region information and Height Functions are insensitive to geometric transformations and nonlinear deformations. The Euclidean distance metric in case of DCT and Dynamic Programming (DP) in case of HF were respectively employed to obtain similarity values and hence fused to classify given query shape based on minimum similarity value. The experiments are conducted on publicly available shape datasets namely MPEG-7, Kimia-99, Kimia-216, Myth and Tools-2D and the results are presented by means of bull’s eye score and precision-recall metric. The comparative study is also provided with the well known approaches to exhibit the retrieval accuracy of the proposed approach. The experimental results demonstrate that the proposed approach yields significant improvement over some of the well known algorithms. ©c 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license  2015 The Authors. Published by Elsevier B.V. (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-reviewunder underresponsibility responsibility organizing committee ofSecond the Second International Symposium on Computer Vision and the Peer-review of of organizing committee of the International Symposium on Computer Vision and the Internet Internet (VisionNet’15). (VisionNet’15)

Keywords: Discrete Cosine Transformation, Height Functions, Dynamic programming, Euclidean distance, Decision fusion, Combined classifier, Shape representation, Shape classification

1. Introduction In the field of computer vision and machine learning, the task of identifying objects in an image or a video sequence is one of the fundamental problems. The identification of objects need a powerful representation, extracting the most discriminating features and also needs a suitable metric to match these features and classifying objects into an appropriate class. In most of the imaging applications, the image analysis can be reduced to the analysis of only shapes, as shape of the object contains perceptual information which is used to describe both object boundary as well as content. Shape based methods may take contour information and/or region information of the shape and it may represent either by extracting local and/or global features. Since object representations exist in different forms, it ∗

Corresponding author E-mail address: [email protected]

1877-0509 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of organizing committee of the Second International Symposium on Computer Vision and the Internet (VisionNet’15) doi:10.1016/j.procs.2015.08.092

715

B.H. Shekar and Bharathi Pilar / Procedia Computer Science 58 (2015) 714 – 722

might be too ambitious to expect a single automated representational system to cope with variety of representations and visual forms from recognition perception. Hence, we are motivated to develop a combined classifier approach that takes advantage of more than one representation scheme and emerge as a powerful classifier system. A brief review of some well known approaches are given here. The Shape Context (SC) 1 captures the spatial distribution of every sample point relative to all other sample point on shape contour. Ling and Jacobs 2 proposed a method called Inner Distance Shape Context (IDSC), where inner distance is used to find the distance between two sample points, for the purpose of shape representation 2 . The contour points distribution histogram (CPDH) 3 descriptor computes the deformable potential at every point along a curve and makes use of EMD distance as the similarity measure for shape matching. The Height Functions proposed by Wang et al., 4 represents the object contour by a fixed number of sample points. Each sample point is associated with a height function. There are methods where multiple features are fused that results in feature level fusion or decision level fusion for the purpose of better retrieval accuracy. One such method is learning manifold approach 5 , in which dissimilarity matrices of two categories of shape descriptors are fused resulting in considerable improvement in retrieval results. For every shape sample, centroid distance, farthest distance, Zernike distance and major axis shape descriptors are obtained. Then the dissimilarity matrices have been fused and the adjacency matrix is formed for the purpose of manifold learning. On the similar line, Local Binary Pattern and Height Functions 6 , Inner distance Shape context and Local Binary Pattern 7 contributions have been explored for shape representation and classification. In the context of designing combined classifier, we have made an attempt to integrate 2D-DCT that capture spatial distribution and Height Functions that captures the contour characteristics and hence to achieve better classification accuracy. The remaining part of the paper is organized as follows. In Section 2, the technique of Discrete Cosine Transformation (DCT) is presented. In Section 3, an insight into Height Functions is given. The proposed approach is given in Section 4. In Section 5, the experimental set-up along with results are brought out and conclusion is given in Section 6.

2. Discrete Cosine Transformation The Discrete cosine transform (DCT) 8 has been widely used in the domain of Image Processing and Pattern Recognition. DCT extracts the frequency domain information from the given object and represents object in terms of few set of coefficients. In DCT coefficients matrix, the first coefficient is called DC coefficient and the other coefficients are referred to as AC coefficients. The frequency of the coefficients increases from left to right and from top to bottom. The DCT coefficients with larger magnitude are mainly concentrated at the upper-left corner which represents the main components of the spectral coefficients of the image. This low frequency part carries most of the visually significant information of the image data. The coefficients at the lower right corner of the matrix represent high frequency part with small amplitude. Since the main energy of the image is concentrated in the low frequency, we can discard some of the AC coefficient values which are close to or equal to 0. For an M x N image, the 2D DCT is given by: C(u, v) = α(u)α(v)

M −1 N −1   x=0 y=0

f (x, y) × cos

π(2y + 1)v π(2x + 1)u cos 2M 2N

(1)

where, ⎧ ⎨ √1 ; u = 0 α(u) = M2 ⎩ M ; u = 1, 2 . . . M − 1

(2)

and, x and y are spatial coordinates in the image, and u and v are coordinates in the transformed image. The 2DDCT coefficients are read in a zigzag manner starting from the top-left corner of the DCT matrix and can be converted to a one dimensional vector. A typical 10x10 DCT matrix transform will have most of the energy relocated to its upper-left corner with the DC coefficient, F00 representing the proportional average of the sample values of the DCT matrix. The AC coefficients of the DCT matrix represents the change in the intensity among the samples.

716

B.H. Shekar and Bharathi Pilar / Procedia Computer Science 58 (2015) 714 – 722

3. Height Functions The Height Functions (HF) is one of the well known contour based shape descriptor proposed by Wang et al., 4 , which is based on the work of Liu et al., 9 . In this approach, a fixed number of sample points on shape contour is considered in order to compute the height functions of a sample point, say x and the height function for x is defined as the distance of other sample points to its tangent line. To illustrate, let p1 , p2 . . . pn be the n sample points on shape contour in a counter clockwise direction. The first step in computing the height functions is to identify a reference axis, which is the tangent line, say li , for every point pi . Let hi,k be the distance between k th sample point to tangent line li . The height functions are signed in nature. The sign is positive, negative or zero depending on whether k th sample point is to the left, right or on the tangent line li , thus providing the information about both the distance and the location of the sample point with respect to the tangent line.

Fig. 1. Height Functions of the jar shape of MPEG-7 dataset

The Figure 1 illustrate the height function for a sample point xi . The height value hi,v is positive since sample point xv is to the left of the axis li and hi,u is negative since the location of xu is to right of the axis. Thus, height values are calculated for every sample points on the contour and are of the same order as the sequence of sample points considered. The height values hi,k (k = 1, . . . , n) of the k th sample point xk with respect to axis li of the point xi is defined as Hi = (h1i , h2i , ..., hni )T = (hi,i , hi,i+1 , ..., hi,n , hi,1 , hi,2 , ..., hi,i−1 )T

(3)

It is observed that h1i =hi,i = 0 for every i = 1, ..., n. The height functions defined above are insensitive to geometric transformations. It is found that HF is sensitive to local contour deformations and hence to overcome this problem smoothing function is designed which is defined as follows. For a given integer k, the sequence of integers 1, 2, . . . , n are divided into disjoint intervals [1, k], [k + 1, 2k], . . ., and the average value of the height values in each interval is computed. fij =

1 k

jk 

hti ,

(4)

t=(j−1)k+1

where j = 1, ..., m with m = n/k and the arithmetic is modulo n. The smoothing process also reduces dimensionality from n to m by the ratio k, since k > 1 and m < n consequently. In order to make shape representation scale invariant, smoothed height values are subjected to normalization by maximal absolute value. The matching of the test shape and training shape is achieved by means of dynamic programming method. The distance between two shapes is computed by finding the optimal correspondence of contour points and the dissimilarity value is the sum of the distances of these corresponding points. This dissimilarity value is used to rank the database of shapes for shape retrieval.

717

B.H. Shekar and Bharathi Pilar / Procedia Computer Science 58 (2015) 714 – 722

4. Proposed Approach In our work, we proposed a combined classifier model that integrates 2D-Discrete Cosine Transformation and Height Functions to represent and classify shapes accurately. The Discrete Cosine Transformation, which captures the spatial information of the shape and the height functions corresponding to sample points of the shape contour are calculated. The Euclidean distance measure is used to compare the DCT values and Dynamic programming (DP) algorithm is used for shape matching in case of height functions. 4.1. Two dimensional Discrete Cosine Transform The feature extraction is carried out using two dimensional discrete cosine transform as follows 1. 2. 3. 4.

For every shape apply 2D-DCT method and obtain the transformed DCT matrix. Take only 10 x 10 matrix of the top left corner of DCT matrix. Convert the above 2D matrix into a 1D array applying zigzag scan Normalize the vector and store in the knowledge base.

Repeat the above process for every shape in the training set forming the DCT based knowledge base. 4.2. Feature extraction based on Height Function The height functions of the given shape are calculated as follows 1. 2. 3. 4. 5.

Extract the contour of the given shape. Select N sample points along the contour in counter-clockwise direction, (N =100 in our experiment). Compute height function vectors for each sample points (Eq. (3)). Obtain the smoothed height function vectors (Eq. (4)). Normalize the smoothed height functions by dividing them with maximal absolute value.

The above process is repeated for all the shapes in the training set to form the height functions based knowledge base. Thus every shape is described by 2D-DCT and HF features forming a hybrid knowledge base. 4.3. Classification Given the query shape, the classification is done as follows. 1. Compute the 2D-DCT matrix of the test shape and obtain its one dimensional vector applying zigzag scan, say Dt . 2. Extract the height functions based feature vectors of the test shape say Ht 3. In case of DCT feature vectors, compute the distance between the test shape with the sample shapes of the training set using Euclidean distance measure. 4. In case of height functions vectors, Dynamic programming (DP) algorithm is used to find out the correspondence between contour sample points of two shapes. Let DT and DH be the distances obtained due to DP and Euclidean distance measure. 5. The distance vectors DT and DH are fused to form the resultant vector DR as follows DR = DT + βDH where, the value β is obtained experimentally, which are variant to the dataset 6. The distance values are arranged in the ascending order of the distance to obtain top shape matches.

(5)

718

B.H. Shekar and Bharathi Pilar / Procedia Computer Science 58 (2015) 714 – 722

5. Experimental Results and Discussions In this section, we present the experimental results on the standard shape datasets namely: MPEG-7, Kimia-99, Kimia-216, Myth and Tools-2D. The performance of the proposed approach is demonstrated through Bull’s eye score 2 . In addition, we have also presented retrieval rate depicting top-n closest matching shape. 5.1. Experimental Results on MPEG-7 The MPEG-7 dataset consists of 1400 shapes from 70 classes with each class consisting of 20 shape samples. The retrieval rate obtained due to proposed approach for each shape class considering top 40 retrievals is shown in Figure 2. The retrieval rate of the proposed approach on MPEG-7 is calculated and the results obtained due to proposed approach along with the results of the state-of-art algorithms is tabulated in Table 1. We notice that the proposed method achieves the better score of 91 percent on MPEG-7 data set. The retrieval rate depicting top 12 closest matching shape is given Table 2. It shall be observed from Figure 2 that more than 50 percent of the classes exhibit almost 100 percent accurate results, where as shapes belong to device classes exhibit little low performance due to high intra-class dissimilarities and high inter class similarities.

Fig. 2. Class-wise retrieval results for MPEG-7 dataset.

In addition, we have also shown in Figure 3 the improvements made by the proposed approach by presenting the top-12 retrieved shapes for a given query shape and compared with the retrieval results of Height Function approach 4 . The graphs of precision and recall for Height Functions 4 alone and proposed approach are given in Fig 4. The proposed approach exhibit better precision and recall rate. 5.2. Experimental Results on Kimia-99 The Kimia-99 shape dataset consists of 9 classes with 11 samples in each class. The 99 shapes belong to 9 different classes. The retrieval rate depicting the top-10 matching shapes is presented in Table 3. In addition, we have also presented the retrieval results on Kimia-99 database considering some of the well known shape representation techniques. It shall be observed from Table 3 that the proposed approach exhibits better performance when compared to other approaches.

719

B.H. Shekar and Bharathi Pilar / Procedia Computer Science 58 (2015) 714 – 722

Fig. 3. Top 12 retrieved shapes of query shape by Height Functions 4 and proposed approach

Table 1. Retrieval rate(Bull’s eye Score) for MPEG-7 dataset 10 - A comparative Analysis

DataSet Proposed Approach HT+PS 11 Height functions 4 Locally affine invariant descriptors 12 Contour flexibility 13 Learned manifold 5 Two strategies 14 Aspect shape context 15 Hierarchical parts 16 Shape tree 17 TAR + shape complexity + global 18 TAR + shape complexity 19 SC + DP 20 HPM 21 Symbolic representation 22 PS+IDSC 23 IDSC + DP 2

MPEG-7 91.00 89.88 89.66 89.62 89.31 88.52 88.39 88.30 88.30 87.70 87.23 87.13 86.80 86.35 85.92 85.82 85.40

Table 2. Top 12 closest matching shapes for MPEG-7 dataset

Dataset PS+LBP 24 IDSC 2 HF 4 HT+PS 11 Proposed Approach

1 1400 1400 1400 1400 1400

2 1345 1375 1384 1384 1386

3 1276 1342 1358 1363 1381

4 1221 1315 1352 1356 1365

5 1157 1256 1302 1305 1327

6 1113 1235 1283 1290 1300

7 1070 1209 1275 1281 1290

8 1023 1188 1251 1265 1279

9 995 1134 1223 1226 1232

10 961 1108 1193 1198 1233

11 933 1048 1169 1172 1201

12 898 1045 1153 1167 1193

720

B.H. Shekar and Bharathi Pilar / Procedia Computer Science 58 (2015) 714 – 722

Fig. 4. Precision-Recall diagram for MPEG-7 dataset

Table 3. Top 10 closest matching shapes for Kimia’s 99 dataset

Approach PS+LBP 24 Learn. manifold 5 Hierarch. Parts 16 IDSC 2 Shape tree 17 Two strategies 14 HF 4 Symbolic Repr. 22 Proposed Approach

1 99 99 99 99 99 99 99 99 99

2 97 99 99 99 99 99 99 99 99

3 97 98 98 99 99 99 99 99 99

4 88 98 98 98 99 98 99 98 99

5 88 98 98 98 99 99 98 99 97

6 86 96 97 97 99 99 99 98 98

7 86 95 96 97 99 99 99 98 99

8 90 89 94 98 97 97 96 95 97

9 80 80 93 94 93 96 95 96 99

10 77 65 82 79 86 84 88 94 96

Total 888 917 954 958 969 969 971 975 982

5.3. Experimental Results on Kimia-216 The Kimia-216 shape dataset consists of 18 classes with 12 samples in each class. The top 11 closest matches obtained due to the proposed methodology is shown in Table 4. In Table 4, the retrieval results on Kimia-216 database for the proposed method and other recent well known methods are placed together. One can notice here too that the proposed approach exhibit better performance when compared to other methods. Table 4. Top 11 closest matching shapes for Kimia-216

Approach SC 1 CPDH+EMD(Eucl) 3 CPDH+EMD(shift) 3 PS+LBP 24 HF 4 IDSC 2 Proposed Approach

1 214 214 215 216 216 216 216

2 209 215 215 209 216 215 216

3 205 209 213 205 216 216 216

4 197 204 205 195 215 214 215

5 191 200 203 195 215 211 214

6 178 193 204 197 212 214 212

7 161 187 190 188 211 210 210

8 144 180 180 180 204 212 210

Table 5. Bull Eye Score(Retrieval Rate) of proposed method for Kimia 99, Kimia 216, MPEG-7 dataset

DataSet Kimia-99 Kimia-216 MPEG-7

Bull eye test 100 98.84 91.00

9 131 168 168 179 200 207 207

10 101 146 154 163 194 204 203

11 78 114 123 152 179 193 195

Total 1809 2030 2070 2079 2278 2312 2314

B.H. Shekar and Bharathi Pilar / Procedia Computer Science 58 (2015) 714 – 722

The overall performance of the proposed approach measured using bull’s eye score on Kimia-99, Kimia-216 and MPEG-7 shape datasets is given in Table 5. 5.4. Experimental Results on Myth dataset Myth data set contains 15 shapes (5 humans, 5 horses and 5 centaurs). We have also presented in Table 6, the average distance within the shape classes and between the shape classes. The experiments exhibit that the combined distance measure suits well to separate the objects belong to different classes. It shall be observed from Table 6 that the intra-class distances are substantially less than the inter-class distances. Table 6. Average distance within classes and the average distance between the classes for the Myth dataset

centaurs horses people

centaurs 2.662553717 5.569424549 12.95656119

horses 5.569424549 4.007757096 14.97722919

people 12.95656119 14.97722919 3.052862898

A comparative analysis of the top-5 retrieval results of Myth database for Height Functions 4 , IDSC 2 and the proposed approach is presented in the Table 7. Table 7. Top 5 closest matching shapes for Myth dataset

Approach IDSC 2 Height Functions 4 Proposed Approach

1st 15 15 15

2nd 15 14 14

3rd 9 12 14

4th 4 10 15

5th 10 9 10

Total 53 60 68

5.5. Experimental Results on Tools dataset The Tools-2D consisting of 35 objects of various instruments (a pair of pliers, knives, scissors of different types). A comparative analysis of the top-5 retrieval results of tools database for height functions and the proposed approach is presented in the Table 8. One can notice here too that the proposed approach classification accuracy is better than the height function alone and also better than IDSC 2 which is found to be one of the best shape representation technique.

Table 8. Top 5 closest matching shapes for Tools-2D dataset

Approach IDSC 2 Height functions 4 Proposed Approach

1st 35 35 35

2nd 26 35 35

3rd 16 33 35

4th 12 30 30

5th 10 24 25

Total 99 157 160

6. Conclusion In this work, we have designed the combined classifier model for binary shapes representation based on Discrete Cosine Transformation and Height Function followed by classification using Euclidean distance and DP as the distance measure. The characteristics of height function such as rotation invariance / insensitivity to noise and occlusion along with DCT’s powerful energy compaction capturing spatial information is found to be suitable choice for shape representation which is demonstrated through extensive experimentation. Experimental results on standard shape databases namely MPEG-7, Kimia-99, Kimia-216, Myth and Tools-2D datasets exhibit the success of the proposed fusion approach.

721

722

B.H. Shekar and Bharathi Pilar / Procedia Computer Science 58 (2015) 714 – 722

Acknowledgements The authors would like to thank the support provided by the Department of Science and Technology, Govt. of India vide Project no. INT/RFBR/P-133 and Russian Foundation for Basic Research, Russia vide Project no. RFBR 12-07-92695.

References 1. Belongie, S., Malik, J., Puzicha, J.. Shape matching and object recognition using shape contexts. IEEE Transactions on Pattern Analysis and Machine Intelligence 2002;24(4):509–522. 2. Ling, H., Jacobs, D.W.. Shape classification using the inner-distance. IEEE Transactions on Pattern Analysis and Machine Intelligence 2007;29(2):286–299. 3. Shu, X., Wu, X.J.. A novel contour descriptor for 2d shape matching and its application to image retrieval. Image and vision Computing 2011;29(4):286–294. 4. Wang, J., Bai, X., You, X., Liu, W., Latecki, L.J.. Shape matching and classification using height functions. Pattern Recognition Letters 2012;33(2):134–143. 5. Chahooki, M.A.Z., Charkari, N.M.. Learning the shape manifold to improve object recognition. Machine vision and applications 2013; 24(1):33–46. 6. Shekar, B.H., Pilar, B.. Shape representation and classification through height functions and local binary pattern-a decision level fusion approach. In: Advances in Computing, Communications and Informatics (ICACCI, 2014 International Conference on. IEEE; 2014, p. 2481– 2488. 7. Shekar, B.H., Pilar, B., Kittler, J.. An unification of inner distance shape context and local binary pattern for shape representation and classification. In: Proceedings of the 2nd International Conference on Perception and Machine Intelligence. ACM; 2015, p. 46–55. 8. Er, M.J., Chen, W., Wu, S.. High-speed face recognition based on discrete cosine transform and rbf neural networks. Neural Networks, IEEE Transactions on 2005;16(3):679–691. 9. Liu, H., Latecki, L.J., Liu, W.. A unified curvature definition for regular, polygonal, and digital planar curves. International Journal of Computer Vision 2008;80(1):104–124. 10. Latecki, L.J., Lakamper, R., Eckhardt, T.. Shape descriptors for non-rigid shapes with a single closed contour. In: IEEE Conference on Computer Vision and Pattern Recognition, 2000. Proceedings; vol. 1. IEEE; 2000, p. 424–429. 11. Shekar, B.H., Pilar, B.. An ensemble of morphology based pattern spectrum and height functions for shape representation and classification. In: Advances in Computing, Communications and Informatics (ICACCI, 2014 International Conference on. IEEE; 2014, p. 2264–2272. 12. Wang, Z., Liang, M.. Locally affine invariant descriptors for shape matching and retrieval. Signal Processing Letters, IEEE 2010;17(9):803– 806. 13. Xu, C., Liu, J., Tang, X.. 2d shape matching by contour flexibility. IEEE Transactions on Pattern Analysis and Machine Intelligence 2009; 31(1):180–186. 14. Temlyakov, A., Munsell, B.C., Waggoner, J.W., Wang, S.. Two perceptually motivated strategies for shape classification. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2010, p. 2289–2296. 15. Ling, H., Yang, X., Latecki, L.J.. Balancing deformability and discriminability for shape matching. In: Proceedings of the 11th European Conference on Computer Vision; ECCV’10. Springer-Verlag; 2010, p. 411–424. 16. Payet, N., Todorovic, S.. Matching hierarchies of deformable shapes. In: Proceedings of the 7th IAPR-TC-15 International Workshop on Graph-Based Representations in Pattern Recognition; GbRPR ’09. Springer-Verlag; 2009, p. 1–10. 17. Felzenszwalb, P.F., Schwartz, J.D.. Hierarchical matching of deformable shapes. In: IEEE Conference on Computer Vision and Pattern Recognition. IEEE; 2007, p. 1–8. 18. Alajlan, N., El Rube, I., Kamel, M.S., Freeman, G.. Shape retrieval using triangle-area representation and dynamic space warping. Pattern Recognition 2007;40(7):1911–1920. 19. Alajlan, N., Kamel, M.S., Freeman, G.H.. Geometry-based image retrieval in binary image databases. IEEE Transactions on Pattern Analysis and Machine Intelligence 2008;30(6):1003–1013. 20. Bai, X., Wang, B., Yao, C., Liu, W., Tu, Z.. Co-transduction for shape retrieval. IEEE Transactions on Image Processing 2012; 21(5):2747–2757. 21. McNeill, G., Vijayakumar, S.. Hierarchical procrustes matching for shape retrieval. In: IEEE Computer Society Conference on Computer Vision and Pattern Recognition; vol. 1. IEEE; 2006, p. 885–894. 22. Daliri, M.R., Torre, V.. Robust symbolic representation for shape recognition and retrieval. Pattern Recognition 2008;41(5):1782–1798. 23. Shekar, B.H., Pilar, B.. A decision level fusion of morphology based pattern spectrum and idsc for shape representation and classification. In: Contemporary Computing and Informatics (IC3I), 2014 International Conference on. IEEE; 2014, p. 908–914. 24. Shekar, B.H., Pilar, B.. Shape representation and classification through pattern spectrum and local binary pattern–a decision level fusion approach. In: Fifth International Conference on Signal and Image Processing (ICSIP). IEEE; 2014, p. 218–224.